(October 19, 2014)

Heisenberg groups over finite fields

Paul Garrett garrett@math.umn.edu http://www.math.umn.edu/egarrett/
[This document is
http://www.math.umn.edu/˜garrett/m/repns/notes 2014-15/05 finite heisenberg ssw.pdf]
1. Heisenberg groups
2. Uniqueness for given non-trivial central character
3. Automorphisms of Heisenberg groups
4. Segal-Shale-Weil/oscillator representations
5. The simplest case: SL2 (Fq )
6. Theta correspondence for O(1) × SL2 (Fq )
7. Theta correspondence for isotropic O(2) × SL2 (Fq )
8. Theta correspondence for anisotropic O(2) × SL2 (Fq )

1. Heisenberg groups
Fix a field k not of characteristic 2.

[1.1] Tangible models A small Heisenberg group over k can be modelled by the group of matrices of the
form [1]  
1 x z
H = { 0 1 y  : x, y, z ∈ k}
0 0 1
Larger Heisenberg groups over k are formed by replacing scalar x, y by vectors:

1 x1 x2 ... xn z
 
 0 1 0 ... 0 y1
  
y1
 
1 x z

.. .. ..

0 0 . 0 . .
xn ) , y =  ...  , z ∈ k}
 
H = { 0 1n y = .  : x = ( x1 ...
  
.. ..

 .. . . 0 yn−1 
0 0 1   yn
0 0 ... 0 1 yn 
0 0 ... 0 0 1

Multiplication is
x0 z0 x + x0 z + z 0 + xy 0
     
1 x z 1 1
0 1 y · 0 1 y0  =  0 1 y + y0 
0 0 1 0 0 1 0 0 1
where the 1-by-1 product xy 0 of 1-by-n and n-by-1 matrices is viewed as a scalar. Inverses are
 −1  
1 x z 1 −x −z + xy
0 1 y = 0 1 −y 
0 0 1 0 0 1

The center Z is small:  

1 0 z
Z = { 0 1 0  : z ∈ k}
0 0 1

[1] Strictly speaking, these are non-reduced Heisenberg groups. Various quotients by subgroups of the center, reduced
Heisenberg groups are also technically useful.

1
 Paul Garrett: Heisenberg groups over finite fields (October 19, 2014)

Commutators are
−1  −1
x0 z0 x0 z0 xy 0 − x0 y
    
1 x z 1 1 x z 1 1 0
0 1 y0 1 y0   0 1 y 0 1 y0  = 0 1 0 
0 0 1 0 0 1 0 0 1 0 0 1 0 0 1

This introduces an alternating form h, i on k 2n :

hx ⊕ y, x0 ⊕ y 0 i = xy 0 − x0 y (with x, x0 row vectors, y, y 0 columns)

The commutator computation shows that H/Z is abelian.

[1.1.1] Remark: There are obvious abstractions and de-coordinatizations of the above. For example, we
can fix a k-vectorspace W , let W ∗ be its dual, and define a group structure on W ∗ ⊕ W ⊕ k by the same
pattern of symbols

(λ, w, z) (λ0 , w0 , z 0 ) = (λ + λ0 , w + w0 , z + z 0 + λw0 ) (with w, w0 ∈ W , λ, λ0 ∈ W ∗ , and z, z 0 ∈ k)

2. Uniqueness for given non-trivial central character

The general assertion is that, given a non-trivial central character of a given Heisenberg group, there is a
unique irreducible with that central character.
Let k be a finite field with q elements, q odd. The cardinality of the nth Heisenberg group H over k is q 2n+1 .

On an irreducible complex representation π of H the center Z acts by a character ω, by Schur’s Lemma.

Since H/Z ≈ k 2n is abelian, the irreducibles with trivial central character ω are one-dimensional, the q 2n
characters of the additive group k 2n .
The irreducibles π with non-trivial central character ω are of greater interest. The restriction of such π to
the (maximal) abelian subgroup
 
1 x z
A = { 0 1 0  ∈ H} ≈ k n ⊕ k}
0 0 1

decomposes as a sum of one-dimensional irreducibles, each of which is ω on Z. The commutation relation

     −1      
1 0 0 1 x 0 1 0 0 1 x 0 1 0 0 1 x −xy
0 1 y 0 1 0 0 1 y = 0 1 y 0 1 −y  =  0 1 0 
0 0 1 0 0 1 0 0 1 0 0 1 0 0 1 0 0 1

shows that when one character ψ of A appears in the restriction of π to A, necessarily

     −1  
 1 0 0 1 x 0 1 0 0  1 x −xy
ψ 0 1 y 0 1 0 0 1 y = ψ0 1 0  = ψ(x) · ω(−xy)
0 0 1 0 0 1 0 0 1 0 0 1

also appears. For non-trivial ω, as y ∈ k n varies, x → ω(−xy) ranges through all characters on k n . That
is, when π has non-trivial central character, all characters of A ≈ k n ⊕ k restricting to ω on Z ≈ k appear
at least once in π restricted to A. There are q n such characters for fixed ω, so the dimension of π with any
non-trivial central character is at least q n .

2
 Paul Garrett: Heisenberg groups over finite fields (October 19, 2014)

Existence of an irreducible with given (non-trivial or not) central character ω follows from existence of the
induced representation IndH A ψ, where ψ is any character on A that restricts to ω on Z. Since #(H/A) = q ,
n

the number of characters ψ of A extending ω, each such induced representation is irreducible. That is, for
non-trivial central character ω and any ψ extending ω to A, IndH A ψ is irreducible.

The finiteness quickly shows there is exactly one irreducible π with given non-trivial central character ω,
since the sum of squares of dimensions of irreducibles is the cardinality of the group. From the inequality
X X
q 2n+1 = |H| = 12 + (dim π)2 ≥ q 2n + (q − 1) · (q n )2 = q 2n+1
one−dim0 l π non−trivial ω

we find that the inequality is an equality. Thus, the dimension of π with non-trivial central character is
exactly q 2 , and each character of A that restricts to ω on Z occurs exactly once. That is, for given non-trivial
central character ω, up to isomorphism there is a unique irreducible π with central character ω.

[2.0.1] Remark: This uniqueness is a trivial analogue of the assertion of the Stone-vonNeumann theorem,
which makes the corresponding assertion over R. Of course, no dimension-counting argument is available
over R.

3. Automorphisms of Heisenberg groups

Given an irreducible π of H, and given an automorphism α of H, the twist π α : h → π(αh) is a representation
of H on the same representation space as that of π. For automorphisms α trivial on the center Z of H, the
twist π α has the same central character ω as π. By the uniqueness of the irreducibles with given non-trivial
central character ω, for automorphism α trivial on Z, π α ≈ π as H-representations.

The group of automorphisms of H fixing Z element-wise is large, as becomes visible with a more coordinate-
independent description of the group. Although the matrix description gives an immediate sense of how the
group behaves, it accidentally makes some misleading distinctions via the x, y coordinates above.
In particular, we show that, for a given field k (not characteristic 2), a 2n-dimensional k-vectorspace V with
a non-degenerate alternating form h, i completely specifies a Heisenberg-type group H = H(V, h, i), and any
linear automorphism of V preserving the alternating form gives an automorphism of H.
First, reconsider a Heisenberg group in coordinates. Because the characteristic is not 2 there is an exponential
map
2
1 x z + xy
      
0 x z 0 x z 0 x z 2
exp  0 0 y  = 1n+2 +  0 0 y  + 21  0 0 y  =  0 1 y 
0 0 0 0 0 0 0 0 0 0 0 1
from the Lie algebra  
0 x z
h = { 0 0 y }
0 0 0
of the Heisenberg group H to the Heisenberg group. The coordinates (x, y, z) on the Lie algebra turn out
to be better than the seemingly-natural coordinates on the group H: we will use notation
1 x z + xy
   
0 x z 2
(x, y, z) = exp  0 0 y  =  0 1 y 
0 0 0 0 0 1
Ignoring the center,
xy x0 y 0 0 0
     xy

1 x 2 1 x0 2 1 x + x0 2 + x 2y + xy 0
(x, y, 0) · (x0 , y 0 , 0) =  0 1 y  · 0 1 0
y  = 0 1 y + y0 
0 0 1 0 0 1 0 0 1

3
 Paul Garrett: Heisenberg groups over finite fields (October 19, 2014)

xy 0 − x0 y
= (x + x0 , y + y 0 , 0) · (0, 0, )
2
That is, in Lie algebra/exponential coordinates, letting v = (x, y, 0) and v 0 = (x0 , y 0 , 0),

hv, v 0 i
v · v 0 = (v + v 0 ) · (0, 0, )
2
Using the (x, y, z) coordinates on theLie algebra h, the Lie bracket
 is
0 x0 z 0 0 x0 z 0
  
0 x z 0 x z
[(x, y, z), (x0 , y 0 z 0 )] =  0 0 y   0 0 y 0  −  0 0 y 0   0 0 y
0 0 0 0 0 0 0 0 0 0 0 0

0 0 xy 0 − x0 y
 

= 0 0 0  = (0, 0, xy 0 − x0 y)
0 0 0

Abstracting this computation, for given k-vectorspace V with non-degenerate alternating form h, i, put a Lie
algebra [2] structure h on V ⊕ k by Lie bracket

[v ⊕ z, v 0 ⊕ z 0 ] = 0 ⊕ hv, v 0 i

In exponential coordinates on H, the exponential map h → H with H ≈ V ⊕ k is notated

exp(v ⊕ z) = v ⊕ z

with Lie group structure on H by

hv, v 0 i
(v ⊕ z) · (v 0 ⊕ z 0 ) = (v + v 0 ) ⊕ (z + z 0 + ) (exponential coordinates in H)
2

In the Lie algebra/exponential coordinates, any k-linear map g : V → V preserving the alternating form
gives an automorphism τg of the Lie algebra h ≈ V ⊕ k and and of the Lie group H ≈ V ⊕ k, by

τg (v ⊕ z) = gv ⊕ z (same expression for both h and H)

4. Segal-Shale-Weil/oscillator representations
The uniqueness of the representation π with fixed non-trivial central character ω of the Heisenberg group
H = V ⊕ k, almost gives a representation of the isometry group Sp(V ) = Sp(V, h, i) of h, i on V .
What literally arises is a projective representation ρ of Sp(V ) on π, meaning that each ρ(g) is ambiguous
by a scalar depending on g, as follows. Let π be an irreducible of H on a complex vectorspace X, with
non-trivial central character ω. The twist π g given by π g (h)(x) = π(τg h)(x) is a representation of H on the
same vectorspace X.
By uniqueness, the twist π g is H-isomorphic to π. [3] By Schur’s lemma, this isomorphism is unique up to
constants. That is, there is an H-isomorphism ρ(g) : (π g , X) → (π, X), unique up to scalar multiples.
[2] In positive characteristic, generally a Lie algebra needs some further structure to behave properly, but in this
simple situation the potential troubles do not appear.
[3] The map (π, X) → (π g , X) by the identity mapping on X is most likely not an H-homomorphism.

4
 Paul Garrett: Heisenberg groups over finite fields (October 19, 2014)

That is, ρ(g) : X → X is a (linear) automorphism of X.

For a subgroup Θ ⊂ Sp(V ), a family of choices {ρ(g)} isomorphism ρ(g) : X → X with the multiplicative
compatibility
ρ(xy) = ρ(x) ◦ ρ(y) (for x, y ∈ Θ)
is a Segal-Shale-Weil/oscillator representation of Θ.
The only possible adjustment of an initial choice of ρ(g)’s, to try to achieve the multiplicative compatibility,
is by constants. There is no advance assurance that adjustment of constants to achieve multiplicative
compatibility is possible. [4]
Over a finite field, we can exhibit a compatible family of choices for g in a subgroup of the full isometry group
Sp(V ). For example, Sp(V ) has subgroups of the form O(Q) × Sp(V 0 ) ⊂ Sp(V ) where Q is a non-degenerate
quadratic space, V 0 is a non-degenerate alternating space, and V ≈ Q ⊗ V 0 . The alternating form on Q ⊗ V 0
is
hx ⊗ v, x0 ⊗ v 0 i = hx, x0 iQ · hv, v 0 iV 0
where h, iQ is the symmetric form on Q and h, iV 0 is the alternating form on V 0 . The inclusion
O(Q) × Sp(V 0 ) ⊂ Sp(V ) is induced by the linear isomorphism Q ⊗ V 0 = V .

[4.1] Canonical intertwinings There is a constructions of a family of copies of π facilitating discussion

of twists π g by g ∈ Sp(V ).
A Lagrangian or maximal totally isotropic subspace W of V is one on which hw1 , w2 i = 0 for all
w1 , w2 ∈ W . For fixed non-trivial central character ω, given a Lagrangian subspace W , extend ω trivially to
W ⊕ Z ≈ exp(W ⊕ Z) by ω(exp W ) = {1} and still denote this extension by ω. Let

IW = IndH
W ⊕Z ω = {C-valued f on H : f (ah) = ω(a) · f (h) for a ∈ exp(W ⊕ Z)}

By uniqueness, the unique (isomorphism class) of irreducible(s) with central character ω is π ≈ IW .

A function f ∈ IW is determined by its values on any Lagrangian subspace complementary to W . [5]

g
For a linear map g : V → V preserving the alternating form, writing (g · f )(h) = f (τg h), the twist π g ≈ IW
is identifiable as

{τg f for f ∈ IW } = {f g on H : τg f (ah) = f (ag hg ) = ω(ag ) · τg f (h) for a ∈ exp(W ⊕ Z)} = IW g

By uniqueness, for fixed non-trivial ω all these twists are H-isomorphic. Conveniently, there is a natural
explicit expression for an intertwining TW →W 0 : IW → IW 0 for any two Lagrangian subspaces W, W 0 :
X
f (eu · h)


TW →W 0 f (h) = (writing eu = exp(u))
u∈W 0 /(W 0 ∩W )

Such a map is the identity for W = W 0 . By Schur, every such map is either 0 or an H-isomorphism. To see
that such an intertwining is non-zero, apply it to the function δW ∈ IW given by

 0 for h 6∈ eW +Z
δW (h) =
ω(z) for h = ew+z = ew ez with w ∈ W and z ∈ Z


[4] Indeed, achieving multiplicative compatibility is not possible for odd n over local fields R or Q , for non-trivial
p
reasons.
[5] Thus, often, one finds the vector space I 2
W identified with L on a fixed complementary Lagrangian subspace,
although this presents some hazards.

5
 Paul Garrett: Heisenberg groups over finite fields (October 19, 2014)

We have
1 X 1 X
δW (eu · h) = ω(eu h)


TW →W 0 δW (h) =
#(W 0 ∩ W) 0
#(W ∩ W 0 )
u∈W u∈W 0 :eu h∈exp(W +Z)

Taking h = 1 ∈ H,
1 X 1 X
δW (eu ) = ω(eu ) = 1


TW →W 0 δW (1) =
#(W ∩ W 0 ) 0
#(W ∩ W 0)
u∈W u∈W 0 :eu ∈exp(W +Z)

Thus, TW →W 0 is non-zero.
Fix Lagrangian W , and as a naive approximation to a Segal-Shale-Weil/oscillator representation ρnf on IW
take ρnf (g) to be the composition
g
ρnf (g) = TW g →W ◦ τg : IW −→ IW = IW g −→ IW

There is no guarantee that this normalizes-by-constants to achieve multiplicativity ρnf (xy) = ρnf (x) ◦ ρnf (y).
In fact, it does not, but, with this as starting point, we can hope to adjust to arrange ρ(xy) = ρ(x) ◦ ρ(y).
Multiplicativity holds up to scalars. Therefore, to determine the (scalar!) discrepancy, if any, it suffices to
examine the behavior of a single function, such as δW , evaluated at a single point, for example 1H .

5. The simplest case: SL2(Fq )

Let Q ≈ Fnq be an n-dimensional non-degenerate quadratic space. Take V = Q ⊗ F2q where F2q has the usual
alternating form
h(x, y), (x0 , y 0 )i = xy 0 − x0 y (for x, y, x0 , y 0 ∈ Fq )
This gives a natural imbedding O(Q) × SL2 (Fq ) ⊂ Sp(V ). Let e1 , e2 be the standard basis of F2q , and
Wi = Fq · ei . Use Lagrangian subspace W = Q ⊗ W2 ⊂ V and W 0 = Q ⊗ W1 , and coordinates (u, v, z) ∈ h
with u ∈ W 0 , v ∈ W , and z ∈ Fq . In these coordinates, G = SL2 (Fq ) acts by matrix multiplication on u, v
and trivially on the center:
   
a b a b
τx (u, v, z) = (u, v, z) · x = (u, v, z) = (ua + vc, ub + vd, z) (for x = ∈ SL2 (Fq ))
c d c d

The standard (upper-triangular) parabolic P ⊂ G stabilizes W2 , and, therefore, W . The Bruhat

decomposition has just two cells: SL2 (Fq ) = P ∪ P wo P . For x ∈ SL2 (Fq ),

 W for x ∈ P
Wx ∩ W =
{0} for x 6∈ P


[5.1] Naive approximation to the Segal-Shale-Weil/oscillator representation Our naive approxima-

tion to a Segal-Shale-Weil/oscillator representation ρnf of G on IW = IndH W ⊕Z ω is
 nf
 ρ (x) = τx (for x ∈ P )

ρnf (x) = TW x →W ◦ τx (for x 6∈ P )



[5.2] The big cell Even though SL(2(Fq ) is finite, there is considerable sense in construing generic
elements of SL2 (Fq ) as lying in the big cell P wo P . The heuristic is that adjusting constants to achieve

6
 Paul Garrett: Heisenberg groups over finite fields (October 19, 2014)

multiplicative compatibility for x, y, xy all in the big cell should suggest correct constant adjustment on the
whole SL2 (Fq ).
 
a b
For x = ∈ SL2 (Fq ) with c 6= 0,
c d
  X  
ρnf (x)f (u, 0, 0) = (TW x →W τx f )(u, 0, 0) = (τx f ) (0, v, 0) · (u, 0, 0)
v∈Fn
q

X  uv  X  uv 
= (τx f ) (0, 0, − ) · (u, v, 0) = ω − · f (ua + vc, ub + vd, 0)
n
2 n
2
v∈Fq v∈Fq

With f = δW the only non-zero summand is where ua + vc = 0, giving v = −ua/c, and

 u2 a 
(TW x →W τx δW )(u, 0, 0) = ω
2c

The value at 1H is
(TW x →W τx δW )(0, 0, 0) = 1

a0 b0
   
a b
For x = and y = with c, c0 , c00 6= 0, let
c d c0 d0

a00 b00 a0 b0
      
a b ∗ ∗
= xy = =
c00 d00 c d c0 d0 ca0 + dc0 cb0 + dd0

Compute the composition of TW x →W ◦ τx and TW y →W ◦ τy on δW evaluated at 1H :

X
(TW x →W τx TW y →W τy δW )(0, 0, 0) = (τx TW y →W τy δW )(0, u, 0)
u∈Fn
q

X   X X  
= (TW y →W τy δW ) (0, u, 0)x = (τy δW ) (0, v, 0) · (0, u, 0)x
u∈Fn
q v∈Fn n
q u∈Fq

X X   X X  
= δW (0, v, 0)y · (0, u, 0)xy = δW (vc0 , vd0 , 0) · (uc00 , ud00 , 0)
v∈Fn n
q u∈Fq v∈Fn n
q u∈Fq

X X  vc0 · ud00 uc00 · vd0 

= δW vc0 + uc00 , vd0 + ud00 , − +
2 2
v∈Fn
q u∈Fq
n

The first argument of δW must be 0 to give a non-zero value, so the sum is over u, v such that vc0 + uc00 = 0.
Replacing v by −uc00 /c0 gives
 −uc00 00 00
uc00 · uc00 d0 
00 uc · ud
X
0
(TW x →W τx TW y →W τy δW )(0, 0, 0) = δW 0, d + ud , −
n
c0 2 2c0
u∈Fq

X  c00 (c0 d00 − c00 d0 ) 

= ω u2 ·
n
c0
u∈Fq

Expanding and simplifying,

c0 d00 − c00 d0 = c0 (cb0 + dd0 ) − (ca0 + dc0 )d0 = c(b0 c0 − a0 d0 ) = −c

7
 Paul Garrett: Heisenberg groups over finite fields (October 19, 2014)

Thus,
X  c00 c 
(TW x →W τx TW y →W τy δW )(0, 0, 0) = ω − u2 · 0
n
c
u∈Fq

Meanwhile, we saw that

(TW xy →W τxy δW )(0, 0, 0) = 1
Thus, the (scalar) discrepancy between ρnf (xy) and ρnf (x) ◦ ρnf (y), for x, y, xy in the big cell, is

ρnf (x) ◦ ρnf (y) = TW x →W τx TW y →W τy

X  c00 c  X  c00 c 
= ω − u2 · 0 · TW xy →W τxy = ω − u2 · 0 · ρnf (xy)
n
c n
c
u∈Fq u∈Fq

We want to adjust all the ρnf (x) by constants to recover multiplicativity.

Since the groundfield is Fq , as opposed to R or Qp , the quadratic exponential sums are nearly characters on
F× ×
q . Specifically, with χ : Fq → {1, −1} by assigning value +1 for non-zero squares and −1 for non-squares,

X X
ω(−u2 · α) = χ(α) · ω(−u2 ) (for α ∈ F×
q )
u∈Fq u∈Fq

In the n-dimensional situation, since quadratic forms are diagonalizable,

X X
ω(−u2 · α) = χ(α)n · ω(−u2 ) (for α ∈ F×
q )
u∈Fn
q u∈Fn
q

Thus, for x, y, xy all in the big cell,

X
ρnf (x) ◦ ρnf (y) = χ(c00 )n · χ(c)n · χ(c0 )n · ω(−u2 ) · ρnf (xy)
u∈F n
q

Rearranging, certainly
X
χ(c)n ρnf (x) ◦ χ(c0 )n ρnf (y) = ω(−u2 ) · χ(c00 )n ρnf (xy)
u∈Fn
q

or
χ(c)n ρnf (x) χ(c0 )n ρnf (y) χ(c00 )n ρnf (xy)
P ◦ P = P
u∈Fnq
ω(−u2 ) u∈Fnq
ω(−u2 ) u∈Fn q
ω(−u2 )

This is the same as

ρnf (x) ρnf (y) ρnf (xy)

◦ =
ω(−u2 · c0 ) ω(−u2 · c00 )
P P P
u∈Fn
q
ω(−u2 · c) u∈Fn
q u∈Fn
q

Thus, we have the desired multiplicative compatibility ρ(x)ρ(y) = ρ(xy) for x, y, xy in the big cell, by taking

ρnf (x) χ(c)n TW x →W ◦ τx

 
TW x →W ◦ τx a b
ρ(x) = P = P = P (x = and c 6= 0)
u∈Fn ω(−u2 · c) u∈ F
2
n ω(−u · c) u∈Fn ω(−u2 ) c d
q q q

8
 Paul Garrett: Heisenberg groups over finite fields (October 19, 2014)

[5.3] The small cell On one hand, our naive normalization ρnf (x) for x in the small cell P is just τx . On
the other hand, x ∈ P can be written as a product of two elements from the big cell P wo P , and in several
ways, for example, x = −wo · wo x. Hopefully, any two such re-expressions agree.
The standard (diagonal) Levi component M of P acts on H by
   
a 0 a 0
τx (u, v, z) = (u, v, z) · x = (u, v, z) = (ua, va−1 , z) (with x = ∈ M)
0 a−1 0 a−1
the unipotent radical N by
   
1 t 1 t
τx (u, v, z) = (u, v, z) · x = (u, v, z) = (u, v + ut, z) (with x = ∈ N)
0 1 0 1
and the long Weyl element by
   
0 −1 0 −1
τwo (u, v, z) = (u, v, z) · wo = (u, v, z) = (v, −u, z) (with wo = )
1 0 1 0

Using the commutation rules, in exponential coordinates the action of the unipotent radical is
 
t 1 t
τx (u, 0, 0) = (u, ut, 0) = (0, ut, u2 ) · (u, 0, 0) (with x = )
2 0 1

where u2 is the value of the quadratic form on u. Thus, on f ∈ IW , up to constants the Levi component
acts by dilation and the unipotent radical by multiplication by a quadratic exponential:
  
a 0
 τx f (u, 0, 0) = f (ua, 0, 0) (for x = )

0 a−1

 
 τx f (u, 0, 0) = ω( 2t u2 ) · f (u, 0, 0)
 1 t
 (for x = )
0 1
Although x → τx is already a representation of P on IW , adjustment of constants is required for compatibility
with the action of the big cell.

The big-cell computations suggest the correct renormalization-by-constants for elements in the small Bruhat
cell P . For example, x = −wo · wo x expresses x ∈ P as a product of elements from the big cell, so we might
hope to prescribe a genuine representation ρ by taking
χ(−1)n TW −wo →W ◦ τ−wo χ(a)n TW wo x →W ◦ τwo x
ρ(x) = ρ(−wo ) ◦ ρ(wo x) = P 2
◦ P 2
(for x ∈ P )
u∈Fn ω(−u )
q u∈Fn ω(−u ) q

By uniqueness, this differs

 by a constant
 from ρnf (x) = τx , and the constant can be ascertained by evaluation
a 0
on δW at 1H : for x = ,
0 a−1
  X  
TW −wo →W ◦ τ−wo ◦ TW wo x →W ◦ τwo x δW (0, 0, 0) = τ−wo ◦ TW wo x →W ◦ τwo x δW (0, u, 0)
u∈Fn
q

X   X X   
= TW wo x →W ◦ τwo x δW (−u, 0, 0) = τwo x δW (0, v, 0) · (−u, 0, 0)
u∈Fn
q v∈Fn n
q u∈Fq

X X   uv  X X uv
= τwo x δW − u, v, = δW (va, ua−1 , )
2 2
v∈Fn n
q u∈Fq v∈Fn n
q u∈Fq

X
= δW (0, ua−1 , 0) = q n = q n · (τx δW )(0, 0, 0)
u∈F n
q

9
 Paul Garrett: Heisenberg groups over finite fields (October 19, 2014)

Replacing the normalizing constants, we are hoping that it is sufficient to take

χ(−1)n χ(a)n
ρ(x) = P 2)
· P 2
· q n · τx (for x ∈ P )
u∈Fn
q
ω(−u u∈ Fn ω(−u )
q

An immediate sensibility check is taking x = 12 , that is, a = 1, which should produce ρ(1) = 1: indeed,
reducing to the one-dimensional case by diagonalizing the quadratic form, it is standard that
X 2 X
ω(−u2 ) = ω(−u2 − v 2 ) = χ(−1) · q
u∈Fq u,v∈Fq

This identity simplifies the presentation of ρ(x) on the Levi component, to

 
n a 0
ρ(x) = χ(a) · τx (for x = )
0 a−1
 
1 t
For x = in the unipotent radical, a similar computation gives
0 1
 
TW −wo →W ◦ τ−wo ◦ TW wo x →W ◦ τwo x δW (0, 0, 0) = q n · (τx δW )(0, 0, 0)

Thus, even more simply than for the Levi component, we are taking
 
1 t
ρ(x) = τx (for x = )
0 1

We see that this ρ restricted to P is a genuine representation on IW , just slightly renormalized from τx itself:
 
a ∗
ρ(x) = χ(a)n · τx (for x = )
0 a−1

Thus, the above heuristic fully suggests the re-normalization of constants to obtain a (hopefully genuine)
representation ρ, given by
 
a ∗

n
χ(a) · τ x (for x = ∈ P)
a−1

0


ρ(x) = n
 
χ(c) ∗ ∗
· TW x →W τx (for x = 6∈ P )


∗
P
u∈F n ω(−u 2) c
q

Thus, with constants adjusted for multiplicative compatibility of ρ on the big cell, ρ is completely determined
on the small cell.

[5.3.1] Remark: Thus, the heuristic suggests a complete specification of re-normalizing constants.
Nevertheless, this is not quite a complete proof that ρ is a genuine representation. As presented, some
other cases must be checked, which we do in the following section. Nevertheless, the point of the present
section is a natural sequence of events completely and correctly specifying ρ.

[5.4] Toward complete certification For a complete proof that ρ is a genuine representation, we must
certify the multiplicative compatibilitly ρ(x) ◦ ρ(y) = ρ(xy) for all cases of x, y, xy in the big or small cells.
With or without the multiplicative compatibility, we do have associativity: ρ(x)(ρ(y)ρ(z)) = (ρ(x)ρ(y))ρ(z),
which can be leveraged into a general argument for multiplicative compatibility, after we know that

ρ(x) ◦ ρ(x−1 ) = ρ(1) = 1IW (x in the big cell)

10
 Paul Garrett: Heisenberg groups over finite fields (October 19, 2014)

To verify the latter, we may as well consider the more general situation of x, y in the big cell, and xy ∈ P .
We start from the point in the computation of ρ(x)nf ◦ ρnf (y) on δW at 1H , where xy had not yet been
assumed in the big cell. Namely, letting c, c0 , c00 be the lower left entries of x, y, xy, we had
  X  c00 c 
ρ(x)nf ◦ ρnf (y) δW (0, 0, 0) = ω − u2 · 0
n
c
u∈Fq

Since c00 = 0 for xy ∈ P ,

 
ρ(x)nf ◦ ρnf (y) δW (0, 0, 0) = q n (for xy ∈ P )

Meanwhile,
   
ρnf (xy)δW (0, 0, 0) = τx yδW (0, 0, 0) = δW (0, 0, 0) = 1 (for xy ∈ P )

Thus,
ρ(x)nf ◦ ρnf (y) = q n · ρnf (xy)
Thus, for xy ∈ P ,

ρnf (x) ρnf (y) q n · ρnf (xy)

ρ(x)ρ(y) = P ◦P 0
= P
ω(−u · c) · u∈Fnq ω(−u2 · c0 )
2 2 2
P
u∈Fn
q
ω(−u · c) u∈Fn
q
ω(−u · c ) u∈Fn
q

Since xy ∈ P , the ρ(xy) renormalized as earlier differs from ρnf (xy) only by χ(a00 )n , where

a00 b00 a0 b0 aa0 + bc0

      
a b ∗
= xy = =
0 (a00 )−1 c d c0 d0 ca0 + dc0 cb0 + dd0

Thus, ca0 + dc0 = 0, which is a0 = −dc0 /c, so

−dc0 (−ad + bc) · c0 c0

a00 = aa0 + bc0 = a · + bc0 = = −
c c c
Thus,
χ(−1) · χ(c)
χ(a00 ) = = χ(−1) · χ(c) · χ(c0 )
χ(c0 )
Thus, as hoped,

q n · ρnf (xy) χ(c)n χ(c0 )n · q n · ρnf (xy)

ρ(x)ρ(y) = P P 0
= P P
u∈Fn
q
ω(−u2 2
· c) · u∈Fnq ω(−u · c ) u∈Fn
q
ω(−u2 ) · u∈Fnq ω(−u2 )

χ(c)n χ(c0 )n · q n · ρnf (xy)

= = χ(a00 )n · ρnf (xy) = ρ(xy) (for x, y 6∈ P and xy ∈ P )
χ(−1)n · q n
In particular, as hoped,
ρ(x) ◦ ρ(x−1 = ρ(1) (for x in the big cell)

[5.5] Remaining certifications We have verified that ρ(x)ρ(y) = ρ(xy) when x, y are in the big cell,
whether xy lies in the big cell or in the small cell. The multiplicative compatibility is also visible for both
x, y ∈ P when they are expressed as products from the big cell, and the outcome is examined in detail.
It remains to treat the case where exactly one of x or y is in the big cell, one in the small cell. This could
be done by repeating the same sort of calculations as above. However, especially as prototype for analogous

11
 Paul Garrett: Heisenberg groups over finite fields (October 19, 2014)

computations for larger groups, there is an easier argument, that plays on the idea that generic elements of
the group are in the big cell, and uses the big-cell computations already done.
Namely, given x, y ∈ SL2 (Fq ), let ε be in the big cell, such that xε and ε−1 y are in the big cell. Because
ρ(x) is unambiguously defined as ρ(x) = ρ(a)ρ(b) for any a, b in the big cell so that x = ab, we have
    
ρ(x) ◦ ρ(y) = ρ(xε)ρ(ε−1 ρ(ε)ρ(ε−1 y) = ρ(xε)(ε−1 y) = ρ (xε)(ε−1 y) = ρ(xy)

This completes the verification that ρ is a genuine representation of SL2 (Fq ).

[5.6] Further comments

Again, the natural appearance of a Segal-Shale-Weil/oscillator representation is as a projective representation,
with natural ambiguity of the constants.
When it is possible to choose constants to achieve multiplicative compatibility on SL2 (k), as opposed to
a metaplectic two-fold cover, it is possible to present a correct description of the representation as a fact-
to-be-verified, without indication of how it could be discovered. This is methodologically unsatisfying, but
does dodge the discussion of Heisenberg groups and Stone-vonNeumann theorems. However, without the
discussion of Heisenberg groups, it is difficult to see that the only obstacle is compatible choice of constants.
That is, in that seemingly simpler context, the issue of multiplicative compatibility must be addressed for
arbitrary vectors in the representation space. This is a feasible computation, but misleadingly portrays the
degree to which the multiplicativity might fail, and is more difficult than the simplified and reasonably
ordered computation above.
For groups Spn (Fq ) with larger n, the technical device of specifying a compatible family of adjustments-by-
constants by writing an arbitrary group element as a product of big-cell elements succeeds, as in the simplest
case above. A slightly generalized Bruhat decomposition not with respect to a minimal parabolic, but with
respect to a parabolic fixing a maximal Lagrangian subspace W (sometimes called a Siegel parabolic), has
cells characterized by dim(W ∩ gW ). The general form of the argument is the same.
Over local fields k such as R and Qp , the behavior of the quadratic exponential sums is only slightly less
congenial. Squares are no longer of index 2 in k × , unlike the finite-field case. Instead, norms from quadratic
field extensions are of index 2, and the corresponding quadratic exponential sums are essentially the norm
residue characters. Thus, the projective form of the representation for SL2 (R) and SL2 (Qp ) can be adjusted
to give a genuine representation only for V = Q ⊗ k 2 with dimk Q even.
Further, over local fields, the proof of the Stone-vonNeumann is correspondingly less trivial, since genuine
analytical issues come into play.

6. Theta correspondence for O(1) × SL2(Fq )

The smallest alternating space Q ⊗ F2q has Q = Fq and the quadratic form is literally squaring: u · u = u2 .
The representation space is IW = IndH W ⊕Z ω with W = Q ⊗ Fq e2 ≈ Fq e2 , which as a complex vector space
is identifiable with L2 (W 0 ) with W 0 = Q ⊗ Fq e1 ≈ Fq e1 . The space IW is q-dimensional.

[6.1] Theta lifts from O(1) to SL2 (Fq ) The group O(1) has two representations, the trivial one, and
the unique non-trivial ±1-valued representation. The action of O(1) on functions on W 0 = Q ⊗ Fq e1 is linear,
by (±1)f (u, 0, 0) = f (±u, 0, 0).
The theta lift of a representation π of O(1) is the π-isotype in IW , as a representation of SL2 (Fq ).
The theta lift of the trivial representation of O(1) to SL2 (Fq ) is the restriction ρ+ of ρ to even functions on
W 0 ≈ Fq , and the theta-lift of the ±1 representation of O(1) is the restriction ρ− of ρ to odd functions on
W 0 ≈ Fq . The restrictions ρ+ and ρ− are q+1 q−1
2 -dimensional and 2 -dimensional, respectively.

12
 Paul Garrett: Heisenberg groups over finite fields (October 19, 2014)

Dimension is sufficient to identify these representations in the classification of irreducibles of SL2 (Fq ): there
are exactly two irreducibles of dimension q+12 of SL2 (Fq ), namely, the irreducible summands of the irregular
principal series I±1 induced from the non-trivial ±1-valued character on P . [6] Thus, ρ+ is one of these.
There is a unique supercuspidal representation of GL2 (Fq ) anomalously decomposing as a direct sum of two
q−1 [7] Thus, ρ is one of these two summands.
2 -dimensional irreducibles when restricted to SL2 (Fq ). −

While happy to know the outcomes, we might want identifications depending less on the finite-dimensionality.

[6.2] Jacquet modules The Jacquet modules JN ρπ are identifiable with N -fixed functions in IW . The
action of N is by multiplication by ω( 12 u2 t) for varying t ∈ Fq and u ∈ W 0 . Thus, the Jacquet module is
identifiable withfunctionsin IW viewed as functions on W 0 supported just on {0}. This is a one-dimensional
a 0
space, and x = ∈ M acts on such functions by
0 a−1

(ρ(x)f )(0, 0, 0) = (τx f )(0, 0, 0) = χ(a) · f (0, 0, 0)

Thus, as M -module, the Jacquet module of ρ+ is JN ρ+ ≈ χ, while JN ρ− ≈ {0}. Thus, the non-supercuspidal
(and non-trivial) part of ρ+ is indeed one of the summands of the (irregular) χth principal series, while ρ−
has no non-supercuspidal part.

[6.3] Whittaker/Gelfand-Graev models From a non-trivial additive character ψ form the corresponding
Whittaker/Gelfand-Graev space
SL2 (Fq )
Wψ = IndN ψ
A natural intertwining Sψ from the whole ρ = ρ+ ⊕ ρ− to the Whittaker space is
 
1 X 1 t
(Sψ f )(x) = ψ(−t) ρ(nx)f (where n = and x ∈ SL2 (Fq ))
q 0 1
t∈Fq

An image Sψ f (x) is a function on H. At x = 1 ∈ SL2 (Fq ), this function is still in IW , and

1 X 1 X

Sψ f (1)(u, 0, 0) = ψ(−t) (τnt f )(u, 0, 0) = ψ(−t) f (u, 0, 0)nt
q q
t∈Fq t∈Fq

1 X 1 X  
= ψ(−t) f (u, ut, 0) = ψ(−t) f (0, ut, 12 u2 t) · (u, 0, 0)
q q
t∈Fq t∈Fq

1 X 1 X
= ψ(−t) ω( 12 u2 t) · f (u, 0, 0) = f (u, 0, 0) · ψ(−t) ω( 21 u2 t)
q q
t∈Fq t∈Fq

This is 0 unless ψ(t) = ω( 12 u2 t), in which case it is f (u, 0, 0). For example, taking ψ(t) = ω( 12 v 2 t) for v ∈ F×
q
guarantees that the intertwining Sψ is not the 0-map.

We recall that both q+1

2 -dimensional summands of the irregular principal series for SL2 (Fq ) have a Whittaker
model for one of the two inequivalent characters, but not for the other.
Similarly, both q−1
2 -dimensional summands of the supercuspidal representation of GL2 (Fq ) that properly
decomposes upon restriction to SL2 (Fq ) have a Whittaker model for one character, but not for the other.
[6] This irregular principal series is the restriction to SL (F ) of a regular (hence, irreducible) principal series of
2 q
GL2 (Fq ). Principal series for GL2 (Fq ) whose two characters whose ratio is not {±1}-valued remain irreducible upon
restriction to SL2 (Fq ).
[7] All the other supercuspidals of GL (F ) remain irreducible when restricted to SL (Q ).
2 q 2 p

13
 Paul Garrett: Heisenberg groups over finite fields (October 19, 2014)

Thus, neither of ρ± has a Whittaker model for additive character inequivalent to ω( 12 t). Without some
further argument, such as dimension-count, this does not quite prove that ρ± is exactly as indicated, but
only that no other irreducibles can appear.

7. Theta correspondence for isotropic O(2) × SL2(Fq )

With q odd, on F2q there are exactly two non-isomorphic non-degenerate quadratic space structures Q,
namely, the (isotropic) (x, y) → xy and, up to constant multiples, the (anisotropic) norm from the unique
quadratic field extension of Fq .
In both cases, O(Q) is a dihedral group, since its index over the special orthogonal group SO(Q) is 2, and

F×

 q (for Q isotropic)
SO(Q) ≈ (where N = NFq2 /Fq )
 {α ∈ F× : N (α) = 1} (for Q anisotropic)
q 2

Thus, the two-dimensional irreducibles π of O(Q), when restricted to SO(Q), are of the form π ≈ π1 ⊕ π1∨ for
a one-dimensional representation π1 of SO(Q) with π1∨ (θ) = π1 (θ−1 ) 6= π1 (θ) for θ ∈ SO(Q). In either case,
given an irreducible π of O(Q), the π-isotype ρπ inside ρ, as a representation of SL2 (Fq ), is the theta-lift of
π.

Here we consider the isotropic case.

[7.1] Jacquet modules The Jacquet module JN ρπ of ρπ can be computed as the ρπ (N )-fixed-points
on IW . Since ρ(N ) acts by multiplication by ω( 21 u2 t), where u2 = (x, y)2 = xy is the quadratic form on
W 0 ≈ F2q , the Jacquet module is the collection of functions f in IW with support on the 0-set of Q, namely,
the union U of the two axes.
There are exactly two O(Q)-orbits on U , namely, the origin (0, 0), and the complement U 0 = U − {(0, 0)} of
the origin. For non-trivial π, the support of SO(Q), π-equivariant f ∈ IW cannot include the origin. Since
SO(Q) is simply transitive on each axis, such f is determined up to a constant on each axis. Indeed, since
SO(Q) includes maps at : (x, y) → (tx, t−1 y),

f (0, t−1 ), 0, 0





f (t, 0), 0, 0 = π1 (at ) · f (1, 0), 0, 0 = π1 (at ) · f (0, 1), 0, 0
 
0 a 0
In particular, as M -representation, with f supported on U , with m = ,
0 a−1

(ρ(m)f )((t, 0), 0, 0) = (τm f )((t, 0), 0, 0) = f ((ta, 0), 0, 0) = π1 (a) · f ((t, 0), 0, 0)

Similarly,

(ρ(m)f )((0, t), 0, 0) = (τm f )((0, t), 0, 0) = f ((0, ta), 0, 0) = π1 (a−1 ) · f ((0, t), 0, 0)

Thus, the Jacquet module of ρπ is two-dimensional:

JN ρπ ≈ π1 ⊕ π1∨

For π1 6≈ π1∨ , this Jacquet module is uniquely identifiable as the Jacquet module of the π1th principal series
representation of SL2 (Fq ). This suggests that ρπ is that principal series, although all that is proven is that
there are no other non-supercuspidal summands.
By direct observation, SO(Q) is transitive on level sets {u ∈ W 0 : u2 = c} for c 6= 0. Thus, the values of f
in the π-isotype is uniquely determined up to a constant on each such level set with c 6= 0, while the values

14
 Paul Garrett: Heisenberg groups over finite fields (October 19, 2014)

on the two axes are independent of each other. Thus, dim ρπ = (q − 1) + 2 = q + 1. This is the dimension
of a principal series, so ρπ is precisely the π1th principal series of SL2 (Fq ).

[7.1.1] Remark: Although it is more difficult to precisely exclude any possibility of supercuspidal summands,
the fashion in which Jacquet modules can be determined is a prototype for the analogous problem over local
fields.

8. Theta correspondence for anisotropic O(Q) × SL2(Fq )

Now consider anisotropic two-dimensional quadratic space Q. Up to a constant multiple, the quadratic form
is the norm from the unique quadratic extension of Fq .

[8.1] Jacquet modules The Jacquet module JN ρπ of the theta lift ρπ of π consists of functions in IW
supported on the isotropic vectors, now just {0} ⊂ W 0 . For non-trivial character π1 of the norm-one subgroup
of F×
q 2 ≈ SO(Q), no function supported on {0} can be SO(Q), π1 -equivariant. Thus, the Jacquet module of
ρπ is {0}, so ρπ is supercuspidal.
For trivial π of O(Q), the Jacquet module of ρπ is the one-dimensional space of functions supported on
isotropic vectors, that is, supported at 0 ∈ W 0 ≈ F2q . The Levi component M acts trivially. The trivial
representation and the special subrepresentation of the principal series induced from the trivial representation
of M both have one-dimensional Jacquet modules with M acting trivially.

[8.2] Dimension of ρπ
We observe that SO(Q) is transitive on {u ∈ W 0 : u2 = c} for each fixed c ∈ Fq . Using the finite-
dimensionality, with π1 6≈ π1∨ , this implies that dim ρπ is at most (q −1)-dimensional, since f in the π-isotype
must vanish for u2 = 0. Thus, ρπ is almost surely a supercuspidal irreducible.
q−1
The only possible complication would be that ρπ were the direct sum of the two anomalous 2 -dimensional
supercuspidal irreducibles. We neglect this point.
To prove that the supercuspidal representations of SL2 (Fq ) thereby produced are distinct, we determine and
compare kernels K(u, v) for ρπ (x), meaning that, with fixed x ∈ SL2 (Fq ),

X
(ρπ (x)f )(u, 0, 0) = K(u, v) f (v, 0, 0)
v∈F2q

For example,

X
trρπ (x) = K(v, v)
v∈F 2
q

 
a b
ω(− 21 t2 ),
P
For x = with c 6= 0, and letting γ = t∈F2q
c d

15
 Paul Garrett: Heisenberg groups over finite fields (October 19, 2014)

1 X  
(ρ(x)f )(u, 0, 0) = (TW x →W τx f )(u, 0, 0) = (τx f ) (0, v, 0) · (u, 0, 0)
γ 2
v∈Fq

1 X −uv 1 X −uv
= (τx f )(u, v, ) = f (ua + vc, ub + vd, )
γ 2
2 γ 2
2
v∈Fq v∈Fq

1 X  (ua + vc)(ub + vd) − uv 

= f (0, ub + vd, ) · (ua + vc, 0, 0)
γ 2
2
v∈Fq

1 X  (ua + vc)(ub + vd) − uv 

= ω · f (ua + vc, 0, 0)
γ 2
2
v∈Fq

Change variables in the sum, replacing v by (v − ua)/c, to obtain

1 X  v(ubc + (v − ua)d) − u(v − ua) 

ω · f (v, 0, 0)
γ 2
2c
v∈Fq

Using ad − bc = 1, this is
1  v 2 d − 2uv + u2 a) 
·ω
γ 2c
This is the kernel for the whole ρ(x). Composing with the projection to the π-isotype, the kernel for ρπ is

1 X  1  (αv)2 d − 2u · αv + u2 a 
π1 (α) + π1∨ (α) ·ω
q+1 γ 2c
α∈SO(Q)

Since α preserves the quadratic form v → v 2 , this is

1 X    v 2 d − 2u · αv + u2 a 
K(u, v) = π1 (α) + π1∨ (α) ω
γ(q + 1) 2c
α∈SO(Q)

1  v 2 d + u2 a  1 X    −u · αv 
= ω · π1 (α) + π1∨ (α) ω
γ 2c q+1 c
α∈SO(Q)

The trace is
X 1 X  v 2 (d + a)  1 X    −v · αv 
trρπ (x) = K(v, v) = ω · π1 (α) + π1∨ (α) ω
γ 2c q+1 c
v∈F2q 2
v∈Fa α∈SO(Q)

The bilinear form attached to the norm form from the quadratic extension is

σ
u·v = 1
2 (u v + v σ u) (non-trivial Galois automorphism σ)

so
 −v · αv   −v 2 (α + ασ ) 
ω = ω
c 2c
We can exploit  2
X q (for t = 0)
ω(t · v 2 ) = (for t ∈ Fq )
−q (for t 6= 0)

v∈F2q

16
 Paul Garrett: Heisenberg groups over finite fields (October 19, 2014)

by judicious choices of x, for example,  

0 −1
x =
1 d
With such x,

1 X  X  v 2 (d − (α + ασ )) 
trρπ (x) = π1 (α) + π1∨ (α) ω
γ(q + 1) c
α∈SO(Q) v∈F2q

Given β ∈ SO(Q), β 6= ±1, take d = β + β σ ,


1 X    q2 (for α = β, β σ )
2q  
trρπ (x) = π1 (α) + π1∨ (α) = π1 (β) + π1∨ (β)
γ(q + 1) γ
−q (for α 6= β, β σ )

α∈SO(Q)

A similar outcome holds for β = ±1, without the 2 in the numerator.

Thus, for another representation π 0 ≈ π2 ⊕ π2∨ of O(Q), with π2 6≈ π1 , π1∨ , for β ∈ SO(Q) such that
π1 (β) 6= π2 (β), π2∨ (β), the traces trρπ and trρπ0 differ at the corresponding x. Thus, the supercuspidal theta
lifts ρπ and ρπ0 are mutually non-isomorphic for two-dimensional irreducibles π and π 0 of O(Q).

17